OpenAI Releases GPT-5.5 Instant as the New Default for ChatGPT
On May 5, OpenAI released GPT-5.5 Instant, which becomes the new default model for ChatGPT. Replacing the previous GPT-3.5 Instant, the model significantly improves the balance between response speed and performance, and it is offered as the default for free users as well.
GPT-5.5 Instant succeeds GPT-3.5 Instant, which free users have relied on until now. In recent years, OpenAI has focused not only on strengthening its expensive flagship models but also on improving the quality of lightweight models that serve as the baseline for everyday use, and this release appears to be part of that effort. The overall ChatGPT user experience is being lifted, which could also affect how it differentiates from competing services.
Google, xAI, and Microsoft Agree to National Security Reviews for New AI Models
According to a report by the Financial Times, Google, xAI, and Microsoft have agreed to submit newly developed AI models to U.S. national security reviews. This signals that the voluntary regulatory framework being worked out between the government and major AI companies is taking concrete shape.
The reviews will target models deemed to be frontier models that exceed certain capability thresholds. Each company will undergo government security evaluations before releasing its models and take risk mitigation measures as necessary. While there has been plenty of debate over AI security risks, the fact that three major providers reached an agreement at the same time suggests that U.S. AI governance policy is starting to gain real effectiveness.
Pennsylvania Sues Character.AI Over Chatbots Impersonating Doctors
On May 5, the state of Pennsylvania filed a lawsuit against Character.AI. According to the state's claims, there have been confirmed cases of Character.AI's chatbots responding to users while posing as medical professionals, which it says threatens public health and safety.
The problem of AI chatbots masquerading as authoritative experts has been pointed out for some time. The risk of users, especially younger ones, taking AI responses as medical advice is real, and regulators' responses have not kept pace. This lawsuit is one of multiple legal actions against Character.AI, indicating that legal and regulatory pressure on AI companies continues to mount.
ElevenLabs Reaches $500 Million in Annual Revenue, With New Investments From BlackRock and Others
ElevenLabs, a leading voice AI company, has disclosed that its annual recurring revenue (ARR) has reached $500 million. At the same time, the company announced that it has brought on BlackRock, Jamie Foxx, and Eva Longoria as new investors.
ElevenLabs has grown rapidly on the back of its text-to-speech and voice cloning technologies, and its enterprise deployment is expanding as well. Voice AI is establishing itself as the next major interface after text-generating AI, and the company's rapid growth reflects how broad that market has become. The entry of institutional investors like BlackRock suggests that the voice AI market is maturing as an investment target.
Meta Uses AI Image Analysis to Detect Minors From Skeletal Structure and Body Size
Meta is stepping up efforts to detect minors through AI-powered image analysis on Instagram and Facebook. The system analyzes skeletal structure and body size in photos to determine the likelihood that a user is a minor.
While the technology is intended to protect minors' safety, it is also raising privacy concerns. Having AI scan and analyze users' physical characteristics has a nature close to biometric identification. Meta describes it as "part of its safety measures," but how much data collection and analysis should be permitted will continue to require debate.
SAP Acquires Dremio to Strengthen Its Agentic AI Foundation
SAP has announced the acquisition of Dremio, a data lake engine company. The goal of the acquisition is to integrate SAP and non-SAP data and strengthen the foundation for agentic AI.
One of the challenges in enterprise AI is whether diverse internal data sources can be leveraged across the organization. SAP already covers a wide range of large-enterprise business systems, but integration with data outside SAP has been a bottleneck. With Dremio's technology, query performance on data lakes is expected to improve, enabling AI agents to access a broader range of internal data.
Gemma 4 MTP Released; 3x Faster Inference on Google TPUs
Google has released a Multi-Token Prediction (MTP) version of Gemma 4. Two models, 31B and 26B (4B active), are available on Hugging Face.
MTP is a technique that predicts multiple tokens at once rather than one at a time, contributing to faster inference. MTP support has been progressing in tools like llama.cpp for some time, and improving inference efficiency for open-source LLMs has become a major trend.
Google also introduced a diffusion-style speculative decoding method that speeds up LLM inference by 3x on TPUs in a developer blog post. Co-optimization of hardware and algorithms to eliminate inference bottlenecks is starting to deliver results at a practical level.
---